perf(polars): rewrite DataFrame serializer with typed column iterators

[idclark]

Closes #29.

The polars DataFrame → line protocol serializer accessed data row-by-row via col.get(row_idx) (per-cell AnyValue dispatch) and allocated an intermediate String per field via format!, giving it the opposite performance character of the Point path. This PR rewrites it on top of the Point path's machinery, in two commits:

1. refactor(point): extract the ryu/itoa/bool/string value writers from FieldValue::write_lp into shared pub(crate) functions. Pure refactor.
2. perf(polars): rewrite dataframe_to_line_protocol:
   • Columns are resolved, downcast to typed ChunkedArray iterators, and name-escaped once before the row loop; iterators advance in lockstep per row (chunk boundaries handled natively — covered by a new test).
   • Values write through the shared ryu/itoa writers into a single output buffer; all-null-field rows are dropped by truncating back to the row start. No more per-line Strings or final join.
   • Exotic dtypes (i128, binary, list, …) keep a per-row AnyValue fallback with unchanged stringification semantics.
   • Timestamp unit conversion is baked into the reader at construction; written via itoa.

Benchmark

New criterion bench (cargo bench --features polars), 10k rows × 25 columns (1 string tag, 20×f64, 4×i64, datetime timestamp) — the wide-numeric parquet-backfill shape from the user report:

	time	throughput
before	25.9 ms	387K rows/s
after	10.2 ms	976K rows/s

~2.5× faster (criterion: −60% ± 2%, p < 0.05). Remaining time is dominated by ryu float formatting, which is inherent to the output.

Semantics

Public API unchanged; all pre-existing tests pass unmodified except to_timestamp's unit test, which was rewritten to pin the same dtype/precision matrix against the new timestamp_reader (the old function became dead code). Two tests added: narrow/unsigned dtype suffixes, and multi-chunk columns.

One deliberate, visible change: floats now format via ryu on the DataFrame path too (matching Point), so extreme magnitudes emit exponent notation (1e30) instead of expanded decimals. Line protocol accepts both.

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

❌ Patch coverage is 85.87786% with 37 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.63%. Comparing base (4db8ce6) to head (e797557).

Files with missing lines	Patch %	Lines
src/write_dataframe.rs	84.23%	35 Missing ⚠️
src/point.rs	95.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #30      +/-   ##
==========================================
- Coverage   83.44%   82.63%   -0.81%     
==========================================
  Files          10       10              
  Lines        1836     1981     +145     
==========================================
+ Hits         1532     1637     +105     
- Misses        304      344      +40     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[alespour]

nice improvements, looks all good to me 🚀